Vibe Coding? Watch it out...

Join us on: Whatsapp

To See All Articles About Technology: Index of Lessons in Technology

Tags: Artificial Intelligence,Generative AI,Large Language Models,

Apple unveils new Mac Studio that can support LLMs with over 600 billion parameters

To See All Articles About Technology: Index of Lessons in Technology

Apple emphasises AI capabilities of the Mac Studio, with the M3 Ultra model supporting large language models (LLMs) with over 600 billion parameters entirely in memory



Apple on Wednesday (March 5, 2025) announced the latest iteration of its Mac Studio, touting it as the most powerful Mac ever made. Equipped with the new M4 Max and M3 Ultra chips, the compact desktop is designed for professionals who require top-tier performance. The new Mac Studio introduces Thunderbolt 5 for faster connectivity, up to 512GB of unified memory, and 16TB of SSD storage—all within a compact and quiet design meant to sit right on a desk.

Apple claims the new Mac Studio provides massive performance gains compared to previous models, making it an ideal choice for users working with AI, video production, and other intensive tasks.

Apple emphasises AI capabilities of the Mac Studio, with the M3 Ultra model supporting large language models (LLMs) with over 600 billion parameters entirely in memory. This makes it a powerhouse for AI workloads, offering high efficiency for machine learning and creative applications. It is also optimised for Apple Intelligence, which enhances productivity and privacy. The new Mac Studio is available for pre-order today, with general availability starting March 12.

Mac Studio with M4 Max

The M4 Max version of Mac Studio is designed for video editors, developers, engineers, and creative professionals. It features a 16-core CPU, a 40-core GPU, and over 500GB/s of memory bandwidth, making it significantly faster than the M1 Max-based Mac Studio. Apple reports that the M4 Max delivers up to 3.5x the performance of the M1 Max model and is 6.1x faster than the most powerful Intel-based 27-inch iMac. With up to 128GB of unified memory, users can handle complex workflows, including large-scale image processing and high-resolution video editing.

Mac Studio with M3 Ultra

For those needing even greater power, the M3 Ultra version of the Mac Studio is the ultimate professional desktop. It boasts an up to 32-core CPU, an 80-core GPU, and 800GB/s of memory bandwidth, delivering nearly twice the performance of M4 Max in multi-threaded workloads. Apple claims that Mac Studio with M3 Ultra is 2.6x faster than its M1 Ultra predecessor and 6.4x faster than the Intel Xeon-based Mac Pro. It supports up to 512GB of unified memory, the highest ever in a personal computer, making it a game-changer for AI, video production, and 3D rendering.

Apple has upgraded the Mac Studio’s connectivity with Thunderbolt 5, offering speeds up to 120 Gb/s, three times faster than the previous generation. This allows for faster external storage, expansion chassis, and multi-display setups. The M3 Ultra model can drive up to eight Pro Display XDRs at full 6K resolution, making it ideal for professionals who require extensive screen real estate. The Mac Studio also includes a 10Gb Ethernet port, HDMI, an SDXC card slot, built-in Wi-Fi, and Bluetooth, providing a well-rounded set of connectivity options.

macOS Sequoia and Apple Intelligence

The new Mac Studio runs on macOS Sequoia, offering features like iPhone Mirroring, enhanced window management, and a redesigned Safari experience. With Apple Intelligence, users can take advantage of AI-powered writing tools, priority notifications, live transcription, and ChatGPT integration in Siri. Apple ensures that privacy remains a top priority, using on-device processing and Private Cloud Compute to handle AI tasks securely.

Mac Studio (2025) Key Specifications

Feature	Mac Studio with M4 Max	Mac Studio with M3 Ultra
CPU	16-core	32-core (24 performance cores)
GPU	40-core	80-core
Neural Engine	3x faster than M1 Max	32-core for AI/ML tasks
Memory	Up to 128GB unified memory	Up to 512GB unified memory
Feature	Mac Studio with M4 Max	Mac Studio with M3 Ultra
Storage	Up to 16TB SSD	Up to 16TB SSD
Memory Bandwidth	500GB/s+	800GB/s+
Connectivity	Thunderbolt 5, 10Gb Ethernet, HDMI, SDXC	Thunderbolt 5, 10Gb Ethernet, HDMI, SDXC
AI Performance	Runs LLMs efficiently	Handles 600B+ parameter models in memory
Display Support	Multiple 6K displays	Up to 8 Pro Display XDRs at 6K resolution

Ref

Tags: Technology,Generative AI,Large Language Models,Artificial Intelligence,

Text-to-SQL Agent

To See All Articles About Technology: Index of Lessons in Technology

Download Code

from langchain_google_genai import ChatGoogleGenerativeAI

import json
with open('../api_keys.json', mode = 'r') as f:
    api_keys = json.load(f)
    
# Initialize the client with your API key
llm = ChatGoogleGenerativeAI(
    model="gemini-2.0-flash",
    google_api_key=api_keys['ashishjain1547']
)

from langgraph.graph import StateGraph, END

from langchain_community.document_loaders import WebBaseLoader
from langchain.chat_models import ChatOpenAI
import os

from typing import TypedDict, Annotated
class AgentState(TypedDict):
    nlq: str
    metadata: dict
    refined_query: str
    further_refined_query: str
    feedback: str
    itr_count: int

import re
def extract_triple_quoted_json(response_text):
    # This pattern looks for a JSON object (starting with '{' and ending with '}')
    # enclosed in either triple double-quotes or triple single-quotes.
    pattern = r'(?:\'\'\'|""")\s*(\{.*?\})\s*(?:\'\'\'|""")'
    match = re.search(pattern, response_text, re.DOTALL)
    if match:
        return match.group(1)
    return None


class ReEngineerQuery:
    def __init__(self, model):
        self.model = model
        
        graph = StateGraph(AgentState)


        graph.add_node("refine_query", self.refine_query)
        graph.add_node("evaluate_reengineered_query", self.evaluate_reengineered_query)
        graph.add_node("further_refine_query", self.further_refine_query)

        # --- Edges ---


        graph.set_entry_point("refine_query")
        graph.add_edge("refine_query", "evaluate_reengineered_query")
        graph.add_edge("evaluate_reengineered_query", "further_refine_query")


        graph.add_conditional_edges(
            "further_refine_query", 
            self.should_continue, 
            {"end": END, "refine_query": "refine_query"}
        )

        # Compile the graph and store references
        self.graph = graph.compile()

        

        
    
    def refine_query(self, state):
        META_PROMPT_TO_REENGINEER_NLQ = """
You are an expert SQLite query generator. Based on the natural language query and the provided table metadata,
please reengineer the query to clearly specify:
- The specific table(s) that should be referenced,
- The calculations or aggregations to perform,
- The structure of the final SQL query.

NLQ: {nlq}

Table Metadata:
{metadata}

Reengineered Query:
"""

        # Build the meta-prompt by substituting the NLQ and metadata.
        prompt = META_PROMPT_TO_REENGINEER_NLQ.format(nlq=state.get("nlq", ""), metadata=state.get("metadata", ""))
        # Invoke the LLM with the prompt.
        response = self.model(prompt)
        # Return the refined query.


        return {"refined_query": response.strip(), "itr_count": state.get("itr_count") + 1}
    
    def evaluate_reengineered_query(self, state):
        EVALUATE_REENGINEERED_QUERY = """
You are an expert SQLite engineer grading an NLQ for correctness, completeness and clarity. \
Generate critique and recommendations for the NLQ so it can be easily converted to an SQL. \
Please evaluate the reengineered query below:
{refined_query}
"""

        prompt = EVALUATE_REENGINEERED_QUERY.format(refined_query=state.get("refined_query", ""))
        response = self.model(prompt)
        return {"feedback": response.strip()}
    
    def further_refine_query(self, state):
        REENGINEER_QUERY_POST_FEEDBACK = """
You are an expert SQLite query generator. Based on the natural language query, the provided table metadata and feedback,
please reengineer the query based on the feedback given to clearly specify: 
- The specific table(s) that should be referenced,
- The calculations or aggregations to perform,
- The structure of the final SQL query.

NLQ: {nlq}

Table Metadata:
{metadata}

Feedback:
{feedback}

Reengineered Query:
"""     # Build the meta-prompt by substituting the NLQ and metadata.
        prompt = REENGINEER_QUERY_POST_FEEDBACK.format(nlq=state.get("nlq", ""), metadata=state.get("metadata", ""),
                                                        feedback=state.get("feedback", ""))
        # Invoke the LLM with the prompt.
        response = self.model(prompt)

        print(response)


        return {"further_refined_query": response.strip(), "itr_count": state.get("itr_count") + 1}
    
    def should_continue(self, state):
        CHECK_CONSISTENCY = """
You are an expert database query evaluator. Your task is to evaluate two queries \
if they are consistent and mean the same thing. One query is the last query and the other is the reengineered query. \

Last Query: {refined_query}
Reengineered Query: {further_refined_query}

Is the reengineered query consistent with the last query? 
Return a JSON response with the key 'answer': 'yes' or 'no'."""
        
        prompt = CHECK_CONSISTENCY.format(refined_query=state.get("refined_query", ""), further_refined_query=state.get("further_refined_query", ""))
        response = self.model(prompt)

        #extract json from the response

        print(response)
        
        json_response = "{" + response.split("{")[1].split("}")[0].strip().replace("'", '"') + "}"

        json_response = json.loads(json_response)
        print(json_response)

    
        if json_response['answer'] == "yes":
            return "end"
        return "refine_query"

class MyGeminiChatModel:
    """
    Minimal wrapper that expects a prompt and returns GPT-3.5 Turbo response text.
    """
    def __init__(self, api_key):
        self.client = ChatGoogleGenerativeAI(
        model="gemini-2.0-flash",
        google_api_key = api_key
    )
        
    def bind_tools(self, tools):
        return self  # For compatibility with how TranslatorCriticApp uses .bind_tools

    def __call__(self, prompt: str) -> str:
        response = llm.invoke(prompt)
        return response.content

model = MyGeminiChatModel(api_keys["ashishjain1547"])

app = ReEngineerQuery(model)



from IPython.display import Image
Image(app.graph.get_graph().draw_png())
<IMG>

with open('tables.json', mode = 'r') as f:
    metadata = json.load(f)

nlq = "Show me the orders from last year."
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})


SELECT * FROM AB_ORDERS WHERE ORDERYEAR = CAST(STRFTIME('%Y', DATE('now', '-1 year')) AS INTEGER)

nlq = "Show all tables"
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})

To show all tables in the SQLite database, you can use the following SQL query:

```sql
SELECT name
FROM sqlite_master
WHERE type='table';
```

This query retrieves the names of all tables present in the SQLite database by querying the `sqlite_master` table where the `type` column is equal to 'table'. Knowing all the tables in the database is essential for database management, schema understanding, and querying purposes. It helps in identifying the available data structures, relationships between tables, and overall database organization. This information is crucial for developers, analysts, and administrators to effectively work with the database and perform various operations such as data retrieval, manipulation, and optimization.

nlq = "What's in orders table?"
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})

SELECT * FROM AB_ORDERS

nlq = "Show me top 5 categories with respect to orders."
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})

SELECT C.CATEGORYNAME, COUNT(O.ORDERID) AS ORDER_COUNT
FROM AB_CATEGORIES C
JOIN AB_PRODUCTS P ON C.CATEGORYID = P.CATEGORYID
JOIN AB_ORDERDETAILS OD ON P.PRODUCTID = OD.PRODUCTID
JOIN AB_ORDERS O ON OD.ORDERID = O.ORDERID
GROUP BY C.CATEGORYNAME
ORDER BY ORDER_COUNT DESC
LIMIT 5;

nlq = "Which areas are dairy products sold?"
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})

SELECT DISTINCT C.CITY AS AREA
FROM AB_CUSTOMERS C
JOIN AB_ORDERS O ON C.CUSTOMERID = O.CUSTOMERID
JOIN AB_ORDERDETAILS OD ON O.ORDERID = OD.ORDERID
JOIN AB_PRODUCTS P ON OD.PRODUCTID = P.PRODUCTID
JOIN AB_CATEGORIES CAT ON P.CATEGORYID = CAT.CATEGORYID
WHERE CAT.CATEGORYNAME = 'Dairy Products';

nlq = "Compare orders from top two cities with respect to total sales."
refined_query = ""
further_refined_query = ""
feedback = ""
itr_count = 0

result = app.graph.invoke({"nlq": nlq, "metadata": metadata, "refined_query": refined_query, "further_refined_query": further_refined_query, "feedback": feedback, "itr_count": itr_count})

-- Query to compare total sales from the top two cities
SELECT 
    c.CITY AS City,
    SUM(p.PRICE * od.QUANTITY) AS TotalSales
FROM 
    AB_CUSTOMERS c
JOIN 
    AB_ORDERS o ON c.CUSTOMERID = o.CUSTOMERID
JOIN 
    AB_ORDERDETAILS od ON o.ORDERID = od.ORDERID
JOIN 
    AB_PRODUCTS p ON od.PRODUCTID = p.PRODUCTID
GROUP BY 
    c.CITY
ORDER BY 
    TotalSales DESC
LIMIT 2;

Tags: Generative AI,Artificial Intelligence,Technology,Database,

OpenAI’s new model o3 is a big leap, stuns tech world

To See All Articles About Management: Index of Management Lessons

OpenAI's new o3 model demonstrates a significant leap in logical reasoning and code-generation capabilities, offering unprecedented productivity for coders. Industry leaders advocate for developers to adapt and leverage AI tools to enhance their skills, noting the irreplaceable value of human intuition, judgment, and innovation in software development.

OpenAI’s new o3 model arrived with little fanfare but quickly set off a surge of conversation in tech circles. Early demonstrations suggest a leap in logical reasoning and code-generation capabilities, far beyond earlier AI tools. If these benchmark results prove accurate, coders may be on the verge of a profound transformation in how they work—an evolution that, while exciting, also raises questions about job displacement and the future of software development.

Many in the industry see o3 as a productivity boon. “These tools drastically improve productivity by significantly reducing the time taken for routine tasks,” says Krishna Prasad Vyakaranam, CTO at Motivity Labs, a part of Magellanic Cloud. “Developing a feature that used to take days can now be completed in minutes. Coders should view this evolution as inevitable and akin to previous technological transitions, such as moving from books to Google and now from Google to advanced AI models.”

Among those intrigued by o3’s performance is Lalitha Duru, VP of CleverTap Labs, who calls it “definitely the biggest leap AI has taken since its inception,” citing its success on benchmarks like ARC-AGI. Though Duru highlights o3’s resource intensity—at times making it costly—she notes that technology often becomes more affordable over time and sees o3 as “a wake-up call signalling the demand for a better class of developers.”

Others emphasise that coders need not fear automation but should adapt, honing the traits that AI cannot replicate. “The capabilities dem onstrated by o3 are undoubtedly impressive and should be seen as a double-edged sword for coders and software developers,” says Atul Rai, co-founder & CEO of Staqu Technologies. He points out that o3’s strongest advantage lies in tackling repetitive tasks. “Rather than fearing obsolescence, coders should embrace this as an opportunity to evolve, leveraging AI tools to augment their capabilities.”

A similar theme emerges in discussions about the human qualities AI lacks. “AI models like o3 still can not fully replicate creative intuition, moral reasoning, and the understanding of ambiguous requirements,” says Manish Jha, chief information officer at Addverb. “Coders and software developers should welcome and adopt this change with enthusiasm and curiosity, but also remain aware of how such models may shape traditional roles.”

For Lakshminarasimman Raghavan, GVP of technology at Publicis Sapient, the continuing importance of human oversight cannot be overstated. “While coders will have powerful assistance from such models, a lot depends on the human-in-the-loop—the programmer—to ensure the code they write is part of software that actually creates value.”

Nearly all these industry experts recommend that developers confront o3’s rise by strengthening the abilities no AI can imitate. “AI can only extrapolate from existing data and patterns,” says Vyakaranam. “It cannot replicate the uniquely human capacity to think about social impact or sense what might be controversial.”

Tags: Layoffs,Generative AI,Artificial Intelligence,

Optimus Is Born - Tesla, August 2021

To see other books: Summaries

    
The friendly robot

Musk's interest in creating a humanoid robot stretched back to the fascination and fear he felt about artificial intelligence. The possibility that someone might create, intentionally or inadvertently, AI that could be harmful to humans led him to start OpenAI in 2014. It also led him to push related endeavors, including self- driving cars, a neural network training supercomputer known as Dojo, and Neuralink chips that could be implanted in brains to create a very intimate symbiotic relationship between humans and machines.

An ultimate expression of safe AI, especially for someone who imbibed sci-fi as a kid, would be creating a humanoid robot, one that could process visual inputs and learn to perform tasks without violating Asimov's law that a robot shall not harm humanity or any human. While OpenAI and Google were focusing on creating text-based chatbots, Musk decided to focus on artificial intelligence systems that operated in the physical world, such as robots and cars. “If you can create a self-driving car, which is a robot on wheels, then you can make a robot on legs as well,” Musk said.

In early 2021, Musk began mentioning at his executive meetings that Tesla should get serious about building a robot, and at one point he played for them a video of the impressive ones that Boston Dynamics were designing. “Humanoid robots are going to happen, like it or not,” he said, “and we should do it so we can guide it in a good direction.” The more he talked about it, the more excited he got.

“This has the potential to be the far biggest thing we ever do, even bigger than a self-driving car,” he told his chief designer, Franz von Holzhausen.

“Once we hear a recurring theme from Elon, we start working on it,” von Holzhausen says. They began meeting in the Tesla design studio in Los Angeles, where the Cybertruck and Robotaxi models were on display. Musk gave the specs: the robot should be about five-foot-eight, with an elfish and androgenous look so it “doesn't feel like it could or would want to hurt you.” Thus was born Optimus, a humanoid robot to be made by the Tesla teams working on self-driving cars.

Musk decided that it should be announced at an event called “AI Day,” which he scheduled for Tesla's Palo Alto headquarters on August 19, 2021.

AI Day

Two days before AI Day, Musk held a prep meeting with the Tesla team virtually from Boca Chica. That day also included a meeting with the Texas Fish and Wildlife Conservation Office to get support for Starship launches, a Tesla finance meeting, a discussion of solar roof finances, a meeting about future launches of civilians, a contentious walk through the tents where Starship was being assembled, an interview for a Netflix documentary, and his second late-night visit to the tract houses where Brian Dow's team was installing solar roofs. After midnight, he got on his plane and headed for Palo Alto.
“It's draining to have to switch between so many issues,” he said when he finally relaxed on the plane. “But there are a lot of problems, and I have to solve them.” So, why was he now leaping into the world of AI and robots? “Because I'm worried about Larry Page,” he said. “I had long conversations with him about AI dangers, but he didn't get it. Now we barely speak.”

When we landed at 4 a.m., he went to a friend's house for a few hours of sleep, then to Tesla's Palo Alto headquarters to meet with the team preparing for the robot announcement. The plan was for an actress to dress up as the robot and come onstage. Musk got excited. “She will do acrobatics!” he declared, as if in a Monty Python sketch. “Can we make her do cool stuff that looks impossible? Like tap dancing with a hat and cane?”
He had a serious point: the robot should seem fun rather than frightening. As if on cue, X started dancing on the conference room table. “The kid has a real good power pack,” his father said. “He gets his software updates by walking around and looking and listening.” That was the goal: a robot that could learn to do tasks by seeing and mimicking humans.

After a few more jokes about hat-and-cane dancing, Musk began drilling down on the final specifications. “Let's make it go five miles per hour, not four, and give it power to lift a bit more weight,” he said. “We overdid making it look gentle.”
When the engineers said that they were planning to have the batteries swapped out when they ran down, Musk vetoed that idea. “Many a fool has gone down the swappable battery path, and it's usually because they have a lousy battery,” he said. “We went down that path with Tesla originally. No swappable pack. Just make the pack bigger so it can operate sixteen hours.”

After the meeting, he stayed behind in the conference room. His neck was hurting from his old Sumo wrestling accident, and he lay on the floor with an ice pack behind his head. “If we're able to produce a general-purpose robot thatcould observe you and learn how to do a task, that would supercharge the economy to a degree that's insane,” he said. “Then we may want to institute universal basic income. Working could become a choice.” Yes, and some would still be maniacally driven to do it.

Musk was in a foul mood at the next day's practice session for AI Day presentations, which would feature not only the unveiling of Optimus but also the advances Tesla was making in self-driving cars. “This is boring,” he kept saying as Milan Kovac, a sensitive Belgian engineer who ran the Autopilot and Optimus software teams, presented very technical slides. “There is too much here that is not cool. This is a recruiting event, and no one will want to join after seeing these fucking slides.”

Kovac, who had not yet mastered the art of deflecting Musk's blasts, walked back to his office and quit, throwing plans for that evening's presentation into disarray. Lars Moravy and Pete Bannon, his more seasoned and battle-hardened supervisors, stopped him as he was about to leave the building. “Let's look at your slides and see how we can fix this,” Moravy said. Kovac mentioned he could use a whiskey, and Bannon found someone in the Autopilot workshop who had some.

They drank two shots, and Kovac calmed down. “I'm going to get through the event,” he promised them. “I'm not going to let my team down.”
With the help of Moravy and Bannon, Kovac cut in half the number of his slides and rehearsed a new speech. “I sucked up my anger and brought the new slides to Elon,” he says. Musk glanced through them and said, “Yep, sure. Okay.”

Kovac got the impression that Musk did not even remember chewing him out.

The disruption caused the presentation that evening to be delayed by an hour.

It was not a very polished event. The sixteen presenters were all male. The only woman was the actress who dressed up as the robot, and she didn't do any fun hat-and-cane dance routines. There were no acrobatics. But in his slightly stuttering monotone, Musk was able to connect Optimus to Tesla's plans for self- driving cars and the Dojo supercomputer. Optimus, he said, would learn to perform tasks without needing line-by-line instructions. Like a human, it would teach itself by observing. That would transform not only our economy, he said, but the way we live.

Ref: Chapter 64, "Elon Musk" by Walter Isaacson

Tags: Book Summary,Technology,Artificial Intelligence,

39 AI Code Tools - The Ultimate Guide in 2024

To See All Articles About Technology: Index of Lessons in Technology

What are the best AI code tools in 2024?

TL;DR - As of September 2024, most programmers achieve the best results by using Cursor with Anthropic Sonnet 3.5 or OpenAI o1.

AI coding tools are becoming standard practice for many developers. And today, you’ll learn which code generators  and tools are the best ones out there for creating high-quality code with the help of artificial intelligence.

Want to learn more? Read on!

Is it possible to code with AI tools?

Yes, it is possible to code with AI tools.  In fact, leveraging AI tools for coding is not only possible, but it can also significantly enhance productivity and accuracy.

AI code is code written by artificial intelligence (AI), often times utilizing large language models (LLMs). These AI programs can write their own programs or translate from one programming language to another. They also perform tasks like offering assistance in auto-generating documentation and finding code snippets faster.

One of the most popular tools is Open AI’s Codex, an AI system that translates natural language to code. Codex powers GitHub Copilot, another popular AI code tool.

OpenAI Codex is capable of interpreting simple commands in natural language and carrying them out for the programmer. This makes it possible to build on top of the existing application with a natural language interface.

As a general-purpose programming model, OpenAI Codex can be applied to almost any programming task. That said, the tool is in beta and so results will vary.

AlphaCode by DeepMind is another tool that is shaking up the industry. Interestingly, this tool outperforms human coders in certain situations. You see, AlphaCode outperformed 45% of programmers in coding competitions with at least 5,000 participants.

However, there are problems with code generators, too. That's why AI coding tools are used to help developers become more productive and efficient, rather than to replace them entirely.

For example, a Stanford-affiliated research team found that engineers who use AI tools are more likely to cause security vulnerabilities in their apps. Plus, questions around copyright are not entirely resolved.

In other words, AI code tools are not yet completely safe to use. That said, the popularity of these tools means that they can’t be overlooked.

What is AI code written in?

AI code is written in languages supported by the AI code generator. For example, OpenAI Codex is most fluent in Python but is also quite capable in several languages, including JavaScript, Ruby, and TypeScript.

Now, let’s take a look at the best code generators out there.

The best AI code generators and AI development tools

What are some effective AI code generators? The most popular ones include OpenAI Codex, Copilot by Github,  ChatGPT by OpenAI as well as open-source models such as Llama 3.

But there are plenty of other tools out there. I’ve listed them here below, including their features, capabilities, and which companies are behind them. Let’s dive in!

Here are the best AI code generators of 2024.

1. OpenAI (ChatGPT, GPT-4, o1)

GPT-4, OpenAI's latest AI model, is a multimodal tool that excels in programming tasks. It understands and explains code, writes new code, and outperforms existing models on Python coding tasks. Despite its ability to handle complex tasks, it has limitations like reasoning errors and potential security vulnerabilities in the code it produces.  

ChatGPT is primarily a user-friendly interface developed by OpenAI that allows you to interact conversationally with advanced language models like GPT-4 and o1-mini. While it's often referred to as a model, ChatGPT is essentially the platform that enables you to generate or debug code and perform other text-based tasks by communicating with these underlying AI models.

Update May 14th: OpenAI just releaded GPT-4o - their new flagship model that’s as smart as GPT-4 Turbo and much more efficient. With 50% reduced pricing and 2x faster latency, it achieves impressive results.

Update September 16th:  o1 is a new series of AI models designed to enhance reasoning by spending more time thinking through problems before responding, excelling in complex tasks in science, coding, and math. OpenAI o1-mini is a faster, more cost-effective model particularly effective at coding, offering an affordable solution for applications that require reasoning but not extensive world knowledge. Both models are now available in ChatGPT and via the API for users to tackle complex problems efficiently.

Price: Free or $20 for GPT Plus

2. Copilot

Copilot uses publicly available code from GitHub repositories so that users can access large datasets and quickly develop accurate code. The tool detects errors in code and recommends changes to it. You can start using GitHub Copilot by installing one of the extensions in your preferred environment.

Price: $10-$19 - GitHub Copilot is free to use for verified students, teachers, and maintainers of popular open source projects.

3. AWS Bedrock

AWS Bedrock is Amazon Web Services' fully managed service that provides developers with access to a variety of powerful foundation models for building and scaling generative AI applications. For programmers, it offers APIs to interact with models like Amazon's Titan and others from leading AI startups, enabling tasks such as code generation, debugging, and text synthesis. While AWS Bedrock simplifies integrating AI into applications, it may have limitations like model accuracy and potential security vulnerabilities in generated code, so developers should exercise caution and perform thorough testing.

Pricing information can be found here

4. AlphaCode

Another AI-based code generator is Google-backed DeepMind’s AlphaCode, which gives developers access to source code from various language libraries. With AlphaCode, developers can leverage thousands of pre-made libraries, helping them connect and use third-party APIs quickly and easily. AlphaCode is not yet available to the public.

Price: No information available

5. Tabnine

Tabnine is an AI code completion tool that utilizes deep learning algorithms to provide the user with intelligent code completion capabilities. Tabnine supports several programming languages such as Java, Python, C++, and more. This tool is open-source and is used by leading tech companies like Facebook and Google.

Price: Paid plans start from $12/month per seat

6. CodeT5

CodeT5 is an open AI code generator that helps developers to create reliable and bug-free code quickly and easily. It is also open-source and provides support for various programming languages such as Java, Python, and JavaScript. CodeT5 also has an online version as well as an offline version for data security.

Price: Free

7. Polycoder

Polycoder is an open-source alternative to OpenAI Codex. It is trained on a 249 GB codebase written in 12 programming languages. With Polycoder, users can generate code for web applications, machine learning, natural language processing and more. It is well-regarded amongst programmers because of its capability of generating code quickly.

Price: Free

8. Deepcode

DeepCode is a cloud-based AI code analysis tool that automatically scans the codebase of a project and identifies potential bugs and vulnerabilities. It offers support for multiple languages such as Java, Python, and JavaScript. DeepCode is well-regarded for its accurate bug detection.

Price: No information available

9. WPCode

WPCode is an AI-driven WordPress code generator created by Isotropic. It supports both developers and non-technical WordPress creators, allowing them to quickly generate high-quality code snippets. CodeWP supports not only HTML and CSS but languages such as Java and Python. It even includes AI assistants to suggest improvements to code snippets.

Price: Starting at $49

10. AskCodi

AskCodi is a code generator that offers a full suite of development tools to help developers build and ship projects faster. With its AI-based code generation, it helps developers write better code and shorter code blocks, with fewer mistakes. AskCodi can be used to develop both web and mobile applications.

Price: Paid plans start from $7.99/month per seat

11. Codiga

Codiga is a static analysis tool that ensures code is secure and efficient. It supports popular languages like JavaScript, Python, Ruby, Kotlin, and more. With Codiga, you can test your code for vulnerabilities and security issues in real time. It also includes an auto-fixer to quickly address any issues in the code.

Price: Paid plans start from $14/month per seat

12. Visual Studio IntelliCode

Visual Studio IntelliCode is an extension of the Visual Studio Code editor created by Microsoft that provides AI-assisted development experiences to improve developer productivity. It offers smarter IntelliSense completions and helps reduce the amount of time developers spend navigating and debugging code.

Price: Starting from $45/month

13. PyCharm

PyCharm is an AI code completion tool from JetBrains which provides developers with intelligent code completion capabilities. This tool supports various programming languages such as Java, Python, and JavaScript. PyCharm is well regarded for its accuracy and can help developers reduce the amount of time spent on coding tasks.

Price: Starting from $24.90/month per seat

14. AIXcoder

AIXcoder is an AI-powered programming pair designed to aid development teams in writing code. It supports languages such as Java, Python, and JavaScript. This tool also offers a range of features such as automated routine tasks, AI-powered code completion, real-time code analysis and error checks while typing.

Price: No information available

15. Ponicode

Ponicode is an AI-powered code assistant designed to help developers optimize their coding workflow. It uses natural language processing and machine learning to generate code from user-defined descriptions. The tool is maintained by CircleCI.

Price: No information available

16. Jedi

Jedi is an open-source option for code completion in AI. It mostly functions as a plugin for editors and IDEs that use Python static analysis tools.

Price: Free

17. Wing Python IDE Pro

Created by Wingware, Wing IDE is a Python-specific software setup that combines the code editing, code navigation, and debugging mechanisms required to Code and Test Software applications. It offers various features such as an intelligent auto-completing Editor, Refactoring, Multi-Selection, and Code Snippets, which make coding much easier and more efficient.

Price: Annual licenses starting at $179/month

18. Smol Developer

Smol is an open-source artificial intelligence agent designed to function as a personal junior developer, capable of generating an entire codebase from your specific product specifications. Unlike traditional, rigid starter templates, Smol can create any kind of application based on your unique requirements. Boasting a codebase that is simple, safe, and small, it offers the perfect blend of ease-of-understanding, customization, and a helpful, harmless, and honest approach to AI development.

Price: Smol is open-source with a MIT License.

19. Cody (Sourcegraph)

Cody (not to be confused with AskCodi), Sourcegraph's AI tool, is a comprehensive coding assistant. It understands your entire codebase, answers queries, and writes code. Beyond guidance, Cody provides detailed code explanations, locates specific components, and identifies potential issues with suggested fixes. Cody works directly in VS code with an extension.

Price: Cody is free for personal use, Sourcegraph starts at $5k/year

20. CodeWhisperer (Amazon)

CodeWhisperer is a tool developed by Amazon. It offers real-time, AI-driven code suggestions and identifies potential open-source code matches for easier review. It even scans for security vulnerabilities, suggesting immediate patches. An added bonus is its commitment to code safety, always aligning with best security practices such as OWASP guidelines.

Price: Free for personal use, $19/month professional use

21. Bard (Google)

Bard can help with programming and software development tasks, including code generation, debugging and code explanation. These capabilities are supported in more than 20 programming languages including C++, Go, Java, Javascript, Python and Typescript. And you can easily export Python code to Google Colab — no copy and paste required. Bard can also assist with writing functions for Google Sheets.

Price: Google Bard is Free

22. Code Llama (Meta)

Code Llama is a set of large language models specialized for coding, built on the Llama 2 platform. It includes different models for various needs: the general-purpose Code Llama, Code Llama - Python for Python-specific tasks, and Code Llama - Instruct for instruction-based coding. These models vary in size (7B, 13B, and 34B parameters) and can handle up to 16k token inputs, with some improvements on up to 100k tokens. The 7B and 13B models also offer content-based infilling.

Code Llama’s training recipes are available on their Github repository - Model weights are also available.

23. Claude 2 & 3, 3.5 (Anthropic)

Claude 3.5 Sonnet is the latest natural language AI model introduced by Anthropic, a firm established by Dario Amodei, formerly of OpenAI. This new iteration is engineered for enhanced input and output lengths and boasts superior performance relative to its earlier version. In an internal agentic coding evaluation, Claude 3.5 Sonnet solved 64% of problems, outperforming Claude 3 Opus which solved 38%. Users can input up to 100K tokens in each prompt, which means that Claude can work over hundreds of pages of technical documentation. The earlier version, Claude 2 scored a 71.2% up from 56.0% on the Codex HumanEval, a Python coding test.

Their evaluation tests the model’s ability to fix a bug or add functionality to an open source codebase, given a natural language description of the desired improvement. When instructed and provided with the relevant tools, Claude 3.5 Sonnet can independently write, edit, and execute code with sophisticated reasoning and troubleshooting capabilities. It handles code translations with ease, making it particularly effective for updating legacy applications and migrating codebases.

A Stability AI Membership is required for commerical application

24. Stable Code 3B

Stability AI's Stable Code 3B, a new 3 billion parameter Large Language Model specialized in code completion, which is 60% smaller yet performs similarly to the larger CodeLLaMA 7b. This model, trained on diverse programming languages and software engineering-specific data, can run in real-time on modern laptops without a GPU. Stable Code 3B is part of Stability AI's Membership program and offers advanced features like Fill in the Middle capabilities and expanded context size, demonstrating state-of-the-art performance in multi-language coding tasks.

A Stability AI Membership (Starting at $20/mo) is required for commercial applications. Free for non-commercial.

25. Replit AI

Replit AI is an innovative code completion tool designed to streamline your coding experience by offering tailored suggestions that align with the context of your current file. As you delve into coding, the tool intuitively presents inline suggestions, enhancing your efficiency and accuracy. Additionally, Replit AI offers advanced features such as the ability to refine suggestions through code comments, the application of prompt engineering for more relevant results, and the flexibility to toggle the code completion feature on or off within the editor settings, ensuring a customized coding environment tailored to your preferences.

Replit AI is available in Replit's Free tier (Limited) and in their Core tier (Advanced Model).  

26. Plandex

Plandex employs persistent agents that tackle extensive tasks spanning numerous files and involving multiple steps. It segments sizable tasks into manageable subtasks, executing each in sequence until the entire task is accomplished. This tool aids in clearing your backlog, navigating new technologies, overcoming obstacles, and reducing the time spent on mundane activities.

Plandex is open-source on Github

27. Meta AI (Meta Lama 3)

Meta has launched Meta AI, powered by the Llama 3 model with 70 billion parameters.  The model positions itself as a powerful asset for improving application functionalities, but it does not match the customization and transparency of more advanced models like GPT-4 Turbo and Claude Opus. The benefits of Meta's approach to open-source AI are multifaceted, including attracting top talent, leveraging community contributions, fostering standardization and lower costs, building goodwill, and aligning with business models that do not rely solely on AI products.  While it is described as "open weight," providing access to the model's weights, it does not include the full toolkit necessary for reproduction. They also co-developed Llama 3 with torchtune, the new PyTorch-native library for easily authoring, fine-tuning, and experimenting with LLMs.

Moreover, Meta is also currently pretraining a 405B parameter model, signaling an ambitious expansion of its AI capabilities. This larger model, set to be released later, promises even more powerful functionalities and potential industry leadership if it surpasses current leaders like GPT-4 and Claude Opus. Such a development could reshape industry standards and perceptions, especially against competitors who guard their models under the guise of safety concerns. This bold move by Meta not only showcases their commitment to advancing AI technology but also challenges the industry's more cautious narratives around the sharing and utilization of AI models, setting new benchmarks for what’s achievable in AI development.

28. MetaGPT

Not to be confused with Meta AI, MetaGPT is a tool that automates the generation of software development outputs such as user stories, competitive analysis, requirements, data structures, APIs, and documents from a single line of input. It integrates roles typically found in a software company—product managers, architects, project managers, and engineers—into its workflow. These roles are executed by large language models (LLMs) following detailed Standard Operating Procedures (SOPs). The core philosophy behind MetaGPT is "Code = SOP(Team)," emphasizing the application of SOPs to organize and direct the work of its LLM teams. This structure aims to mimic the entire process of a software company, simplifying and automating complex tasks.

MetaGPT is MIT licensed and open-source

29. AutoRegex

AutoRegex is my favorite tool to translate natural language to regex. If you're like me, you wiped all traces of regex syntax from your memory the moment ChatGPT released - this helps!

30. llama.cpp

Llama.cpp is designed to facilitate LLM inference with optimal performance and minimal initial setup across various hardware, both locally and in the cloud. It is implemented in plain C/C++ without dependencies and features extensive support for Apple silicon through ARM NEON, Accelerate, and Metal frameworks. It also supports AVX, AVX2, and AVX512 for x86 architectures and offers integer quantization from 1.5 to 8 bits to enhance inference speed and reduce memory consumption. For NVIDIA GPUs, llama.cpp includes custom CUDA kernels, with AMD GPU support through HIP. Additionally, it supports Vulkan, SYCL, and partial OpenCL backends and can perform hybrid CPU+GPU inference to manage models that exceed VRAM capacity.

31. Aider

Aider is a  command line tool  allowing you to pair program with LLMs directly in your terminal. It seamlessly integrates with your local git repository, editing code directly in your source files and crafting smart commit messages for each change.

Aider is open-source on Github

32. Codestral (Mistral)

A model fluent in 80+ programming languages, Codestral, is Mistrral's first-ever code model. Codestral is an open-weight generative AI model explicitly designed for code generation tasks. It helps developers write and interact with code through a shared instruction and completion API endpoint. As it masters code and English, it can be used to design advanced AI applications for software developers.

Codestral is a 22B open-weight model licensed under the new Mistral AI Non-Production License, which means that you can use it for research and testing purposes. Codestral can be downloaded on HuggingFace

Update July 16th: Codestral Mamba release:  For easy testing, they made Codestral Mamba available on la Plateforme (codestral-mamba-2407), alongside its big sister, Codestral 22B. While Codestral Mamba is available under the Apache 2.0 license, Codestral 22B is available under a commercial license for self-deployment or a community license for testing purposes.

33. Cursor

Cursor is an AI-enhanced code editor designed to boost productivity by enabling developers to interact with their codebase through conversational AI and natural language commands. It includes features like Copilot++, which predicts your next code edit, and Cmd-K, which allows code modifications through simple prompts.

You can try Cursor for free

34. Warp

Warp is a modern, Rust-based terminal with AI built in. Type ‘#’ on your command line and start describing the command you want to run using natural language. Warp will load AI Command Suggestions as you type.

Warp AI is free to use up to 40 requests per user per month. You can create a Team and upgrade to a Team plan to unlock higher Warp AI request limits. Visit the pricing page to learn more.

35. CodiumAI

CodiumAI is a trending tool that developers can use to enhance their coding experience with the power of AI.

Key features: When compared to the other tools, CodiumAI provides a set of unique features:

Precise code suggestions: CodiumAI thoroughly analyzes your code, providing tailored suggestions. These include adding docstrings, refining exception handling, and implementing best practices, directly improving your code’s quality.
Code explanation: This tool offers detailed descriptions of your source code or snippets, breaking down each component and offering insights and sample usage scenarios to enhance code comprehension.
Automated test generation: Testing is essential in large codebases. CodiumAI simplifies this by swiftly generating accurate and reliable unit tests without manual intervention, saving significant time and effort and ensuring thorough testing of your codebase.
Code behavior coverage: Comprehensive testing means covering all possible code behaviors. CodiumAI’s “Behavior Coverage” feature generates test cases covering various code behaviors and seamlessly applies related changes to your source code.
Streamlined collaboration: CodiumAI facilitates teamwork by enabling seamless collaboration among developers. Its Git platform integration allows for sharing and reviewing code suggestions and test cases within your development team, promoting efficient workflows and code quality.
Seamless implementation: With CodiumAI’s intelligent auto-completion agent, implementation becomes effortless. It seamlessly integrates with your task plans, ensuring smooth execution from concept to completion of your code.
Multiple language and IDE support: CodiumAI supports popular programming languages such as Python, JavaScript, and TypeScript while seamlessly integrating with leading IDEs, including VSCode, WebStorm, IntelliJ IDEA, CLion, PyCharm, and JetBrains.
Pricing

The pricing of CodiumAI offers free code integrity for developers at $0/user per month, while teams can access optimized collaboration for $19/user per month.

36. MutableAI

MutableAI is a tool that revolutionizes the coding experience with features such as AI autocomplete, one-click production code enhancements, prompt-driven development, test generation, and extensive language and IDE integration, empowering developers to write code more efficiently and effectively.

Key features

Here are the key features of MutableAI:

AI Autocomplete: Minimize time spent on boilerplate code and searching for solutions on Stack Overflow with specialized neural networks providing intelligent code suggestions.
Production Quality Code: Refactor, document, and add types to your code effortlessly, ensuring high-quality code output.
Prompt-driven Development: Interact directly with the AI by giving instructions to modify your code, enabling a more intuitive and interactive coding experience.
Test Generation: Automatically generate unit tests using AI and metaprogramming techniques, ensuring comprehensive test coverage for your code.
Language and IDE Integration: Supports popular languages like Python, Go, JavaScript, TypeScript, Rust, Solidity, and more, as well as integration with IDEs like JetBrains and Visual Studio (VS) Code.
Pricing

MutableAI’s basic plan offers $2 per repo per month, while its premium plan offers $15 per repo per month.

37. Figstack

Figstack is an innovative AI tool that provides developers with various features to improve code understanding, translation, documentation, and optimization. Figstack caters to developers at all levels, from beginners looking to understand complex code to experienced professionals aiming to automate tedious tasks like writing documentation or measuring code efficiency.

Key features

Code explanation in natural language: This feature helps users easily understand the code written in any language by translating it into clear, natural language descriptions.
Cross-Language code translation: Developers can easily convert code from one programming language to another. This simplifies the process of porting applications across different technology stacks.
Automated function documentation: Figstack automatically generates detailed docstrings that describe the function’s purpose, parameters, and return values, ensuring that your code is always readable, maintainable, and well-documented.
Time complexity analysis: The tool helps developers assess the efficiency of their code in Big O notation, pinpoint bottlenecks, and optimize their code for better performance by identifying the time complexity of a program.
Pricing

Figstack is free to use and includes most of the essential features.

38. CodeGeeX

CodeGeeX is an AI-powered code generation tool designed to assist developers in writing, completing, and optimizing code more efficiently. It leverages deep learning models trained on a wide variety of programming languages and codebases, where it can provide context-aware code suggestions, complete code snippets, and even generate entire functions or modules.

Key features

Code generation and completion: CodeGeeX offers accurate code generation capabilities based on natural language descriptions. Also, it can complete the current line or multiple lines ahead, making the development process faster.
Code translation: Developers can effortlessly convert their code from one programming language to another.
Automated comment generation: The tool saves time by automatically generating line-level comments, which helps improve code readability and maintainability.
AI chatbot: The AI chatbot in CodeGeeX provides quick answers to technical questions directly within the development environment instead of having developers find solutions on the internet.
Wide IDE and language support: CodeGeeX supports various popular IDEs, including Visual Studio Code, JetBrains IDEs and multiple programming languages, such as Python, C++, JavaScript, and Go.
Pricing

CodeGeeX offers their plugin completely free for individual users. If there are more advanced requirements, they provide an enterprise plan.

39. Codeium

One I personally use.

Millions of engineers, including our own, use these features every single day.
Autocomplete
Autocomplete faster than thought. Codeium's generative code can save you time and help you ship products faster.

Command
Give instructions in your editor to perform inline refactors, whether it is generating code, adding comments, or something even more complex.

Chat
Generate boilerplate, refactor code, add documentation, explain code, suggest bug fixes, and so much more. Powered by the largest models, optimized for coding workflows and Codeium's industry-leading reasoning engine.

Context
All of Codeium's features are powered by an industry-leading context awareness and reasoning engine. With full repository and multi repository codebase awareness, Codeium provides 35% more value purely from providing more grounded results. 

References

• code submit
• codium.ai

Tags: Technology,Artificial Intelligence,Generative AI,Large Language Models,Python,JavaScript,

Cisco’s second layoff of 2024 affects thousands of employees

To See All Articles About Layoffs / Management: Index of Layoff Reports

U.S. tech giant Cisco has let go of thousands of employees following its second layoff of 2024. 

The technology and networking company announced in August that it would reduce its headcount by 7%, or around 5,600 employees, following an earlier layoff in February, in which the company let go of about 4,000 employees. 

As TechCrunch previously reported, Cisco employees said that the company refused to say who was affected by the layoffs until September 16. Cisco did not give a reason for the month-long delay in notifying affected staff. One employee told TechCrunch at the time that Cisco’s workplace had become the “most toxic environment” they had worked in.

TechCrunch has learned that the layoffs also affect Talos Security, the company’s threat intelligence and security research unit. 

Cisco said in its August statement that its second layoff of the year would allow the company to “invest in key growth opportunities and drive more efficiencies.” On the same day, Cisco published its most recent full-year earnings report, in which the company said 2024 was its “second strongest year on record,” citing close to $54 billion in annual revenue.

Cisco chief executive Chuck Robbins made close to $32 million in total executive compensation during 2023, according to the company’s filings.

When reached by email, Cisco spokesperson Lindsay Ciulla did not provide comment, or say if Cisco’s executive leadership team planned to reduce their compensation packages following the layoffs.

Are you affected by the Cisco layoffs? Get in touch. You can contact this reporter on Signal and WhatsApp at +1 646-755-8849, or by email. You can send files and documents via SecureDrop.

A look at Cisco’s response to the current economic climate and transition trajectory leading to significant layoffs:

Cisco’s focus on subscription-based services

Cisco's $28 billion acquisition of Splunk in March signals a strategic shift towards subscription-based services. This move marked a significant shift for Cisco, traditionally known for networking equipment, as it entered the competitive cybersecurity market alongside players like Palo Alto Networks, Check Point, CrowdStrike, and Microsoft, as ET followed this development.

Cisco’s funding to AI startups

Since 2018, Cisco has been actively involved in the AI space, acquiring Accompany and CloudCherry to expand its presence in this rapidly growing technology. In 2019, the company launched the Silicon One ASIC chip, offering speeds of 25.6 Tbit/s, directly competing with Intel and Nvidia.

Cisco has allocated $1 billion to fund AI startups.
Earlier in February, Cisco partnered with Nvidia. The latter agreed to use Cisco's ethernet with its own technology that is widely used in data centers and AI applications. In June, Cisco invested in AI startups like Cohere, Mistral AI, and Scale AI. The company announced that it had made 20 acquisitions and investments related to AI in recent years.

Focus on emerging technologies

Cisco offers data center technologies like the Unified Computing System (UCS) and Nexus switches, designed to support modern data center and cloud environments. Additionally, their collaboration tools, such as WebEx and Cisco Jabber, enhance communication and productivity.

Shifting focus on cybersecurity

Since 2013, with the acquisition of Sourcefire, a network security and threat detection provider Cisco strengthened its security portfolio. Open DNS acquired in 2015, provides cloud based threat detection and prevention. CloudLock, a cloud security solutions provider for $293 million protects users and data in cloud environments. Duo Security, for $2.35 billion, provides cloud based authentication and access control.

References

• TechCrunch
• economic times

Tags: Technology,Layoffs,Management,Artificial Intelligence,

How to use AI for coding the right way

To See All Articles About Technology: Index of Lessons in Technology